Skip to content

visualization.md

Latest commit

 

History

History
148 lines (101 loc) · 4.43 KB

visualization.md

File metadata and controls

148 lines (101 loc) · 4.43 KB
Raw

Visualization

Light LLM Simulator produces visualization assets from generated CSV results and exposes them in two ways:

The visualizations are static PNG images generated from simulator output. There is no interactive charting layer in the current implementation.

Output Layout

Generated images are written under data/images/:

  • data/images/throughput/
  • data/images/pipeline/mbn2/
  • data/images/pipeline/mbn3/
  • data/images/pipeline/deepep/

Generated CSV inputs are read from:

  • data/afd/mbn2/best/
  • data/afd/mbn3/best/
  • data/deepep/

CSV filenames follow:

{DeviceType.name}-{ModelType.name}-tpot{tpot}-kv_len{kv_len}.csv

Example:

ASCENDA3_Pod-DEEPSEEK_V3-tpot50-kv_len4096.csv

CLI Tools

Throughput Charts

Script: src/visualization/throughput.py

This script generates two families of throughput charts:

  • throughput vs. total dies for a specific (device, model, tpot, kv_len)
  • AFD-over-DeepEP improvement vs. kv_len and tpot for a fixed total_die

Example

python src/visualization/throughput.py \
  --model_type deepseek-ai/DeepSeek-V3 \
  --device_type Ascend_A3Pod \
  --tpot_list 20 50 70 100 150 \
  --kv_len_list 2048 4096 8192 16384 131072 \
  --micro_batch_num 2 3 \
  --total_die 128 \
  --min_die 16 \
  --max_die 768

Output

  • data/images/throughput/{DeviceType.name}-{ModelType.name}-mbn{micro_batch_num}-total_die{total_die}.png
  • data/images/throughput/{DeviceType.name}-{ModelType.name}-tpot{tpot}-kv_len{kv_len}.png

Notes

  • throughput_vs_dies() expects all three CSVs to exist for the same file name:
    • data/deepep/
    • data/afd/mbn2/best/
    • data/afd/mbn3/best/
  • If one of those files is missing, the script raises FileNotFoundError for that chart.

Pipeline Charts

Script: src/visualization/pipeline.py

This script generates Gantt-style pipeline charts for the available serving modes by reading one CSV filename across:

  • data/deepep/
  • data/afd/mbn2/best/
  • data/afd/mbn3/best/

Example

python src/visualization/pipeline.py \
  --file_name ASCENDA3_Pod-DEEPSEEK_V3-tpot50-kv_len4096.csv

Output

  • data/images/pipeline/deepep/{file_stem}-total_die{total_die}.png
  • data/images/pipeline/mbn2/{file_stem}-total_die{total_die}.png
  • data/images/pipeline/mbn3/{file_stem}-total_die{total_die}.png

Notes

  • The script skips missing CSV inputs per serving mode instead of failing the whole run.
  • One image is produced per matching total_die row in the source CSV.

Web Visualizations Tab

The browser UI includes a Visualizations tab implemented in ThroughputCharts.vue.

How it works

  1. The user selects or seeds:
    • device_type
    • model_type
    • total_die
    • tpot
    • kv_len
  2. The frontend calls:
GET /api/results
  1. The backend returns lists of image URLs under /data/images/...
  2. The UI renders those PNGs directly

The tab is seeded from shared app state, so using View Charts from the Results table carries the first selected row's total_die into the visualization view.

Important behavior

  • The UI is static-image based only. There is no zoom, pan, or export flow.
  • webapp/backend/main.py filters /api/results to return only image files that actually exist on disk.
  • Some parameter combinations legitimately return fewer images than others if only part of the visualization set has been generated.

End-to-End Flow

Typical visualization workflow:

  1. Run a simulation from the CLI or web UI
  2. Confirm CSV outputs exist under data/afd/... or data/deepep/
  3. Generate images with:
  4. Open the web UI and use the Visualizations tab, or inspect the PNG files directly

Limitations

  • Visualizations are pre-generated assets, not live chart computations
  • Missing CSV inputs lead to missing charts
  • Some chart families depend on both AFD and DeepEP results existing for the same parameter set
  • The web UI does not offer chart export beyond the static PNG files already saved under data/images/